Semantic Analysis of Social Media Posts Using a Bidirectional LSTM: A Deep Learning Baseline for Social Impact Theory Based Optimization

Abstract

Social media platforms generate millions of short text posts every day, and these posts carry the opinions, moods, and intentions of their authors. Reading this meaning automatically is the task of semantic analysis, and it has become a basic need for marketing research, public health monitoring, and the study of online behavior. The present work reports the first experimental phase of larger research program whose goal is to combine semantic analysis with Social Impact Theory based optimization. Before social influence signals can be added to a model, a strong and well-understood content-only baseline is required. This paper builds that baseline. A stacked bidirectional Long Short-Term Memory (BiLSTM) network with fine-tuned GloVe word embeddings was trained on a balanced sample of 200,000 tweets from the Sentiment140 collection. The study pays special attention to preprocessing, an area where small mistakes are common and costly. Three findings are documented: the order of the cleaning steps decides whether URL and mention patterns can be matched at all; negation words must survive stop word removal because they flip sentiment; and stemming should not be combined with pre-trained embeddings, since word stems fall outside the embedding vocabulary. The final model reached 79.22% test accuracy, an F1-score of 0.799, and an area under the ROC curve of 0.878. An error analysis shows that the remaining mistakes concentrate on short, ironic, and context-dependent posts, which content-only models cannot resolve. This result gives direct empirical support to the central hypothesis of the research program: further progress requires social context, and Social Impact Theory offers a principled way to bring it into the learning process.